XXX: Use correct encoding for Vertex Fetch instructions on Cayman.cayman-only-bfgminer

3 files changed, 275 insertions, 139 deletions

diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
index 4d6c25c190b..d39e0791fb6 100644
--- a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -99,7 +99,7 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
   } else if (IS_VTX(Desc)) {
     uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
     uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
-    InstWord2 |= 1 << 19;
+//    InstWord2 |= 1 << 19;
 
     Emit(InstWord01, OS);
     Emit(InstWord2, OS);
diff --git a/lib/Target/R600/R600InstrFormats.td b/lib/Target/R600/R600InstrFormats.td
index 618004fd269..1aacda37125 100644
--- a/lib/Target/R600/R600InstrFormats.td
+++ b/lib/Target/R600/R600InstrFormats.td
@@ -173,7 +173,6 @@ class VTX_WORD0 {
   bits<8> BUFFER_ID;
   bits<1> SRC_REL;
   bits<2> SRC_SEL_X;
-  bits<6> MEGA_FETCH_COUNT;
 
   let Word0{4-0}   = VC_INST;
   let Word0{6-5}   = FETCH_TYPE;
@@ -182,9 +181,28 @@ class VTX_WORD0 {
   let Word0{22-16} = SRC_GPR;
   let Word0{23}    = SRC_REL;
   let Word0{25-24} = SRC_SEL_X;
+}
+
+class VTX_WORD0_eg : VTX_WORD0 {
+
+  bits<6> MEGA_FETCH_COUNT;
+
   let Word0{31-26} = MEGA_FETCH_COUNT;
 }
 
+class VTX_WORD0_cm : VTX_WORD0 {
+
+  bits<2> SRC_SEL_Y;
+  bits<2> STRUCTURED_READ;
+  bits<1> LDS_REQ;
+  bits<1> COALESCED_READ;
+
+  let Word0{27-26} = SRC_SEL_Y;
+  let Word0{29-28} = STRUCTURED_READ;
+  let Word0{30}    = LDS_REQ;
+  let Word0{31}    = COALESCED_READ;
+}
+
 class VTX_WORD1_GPR {
   field bits<32> Word1;
   bits<7> DST_GPR;
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index 73767f0ea5e..d6f38cd7f7e 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -261,6 +261,50 @@ class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> mask, dag outs,
 
 }
 
+class VTX_READ <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+    : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
+      VTX_WORD1_GPR {
+
+  // Static fields
+  let DST_REL = 0;
+  // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
+  // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
+  // however, based on my testing if USE_CONST_FIELDS is set, then all
+  // these fields need to be set to 0.
+  let USE_CONST_FIELDS = 0;
+  let NUM_FORMAT_ALL = 1;
+  let FORMAT_COMP_ALL = 0;
+  let SRF_MODE_ALL = 0;
+
+  let Inst{63-32} = Word1;
+  // LLVM can only encode 64-bit instructions, so these fields are manually
+  // encoded in R600CodeEmitter
+  //
+  // bits<16> OFFSET;
+  // bits<2>  ENDIAN_SWAP = 0;
+  // bits<1>  CONST_BUF_NO_STRIDE = 0;
+  // bits<1>  MEGA_FETCH = 0;
+  // bits<1>  ALT_CONST = 0;
+  // bits<2>  BUFFER_INDEX_MODE = 0;
+
+  // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
+  // is done in R600CodeEmitter
+  //
+  // Inst{79-64} = OFFSET;
+  // Inst{81-80} = ENDIAN_SWAP;
+  // Inst{82}    = CONST_BUF_NO_STRIDE;
+  // Inst{83}    = MEGA_FETCH;
+  // Inst{84}    = ALT_CONST;
+  // Inst{86-85} = BUFFER_INDEX_MODE;
+  // Inst{95-86} = 0; Reserved
+
+  // VTX_WORD3 (Padding)
+  //
+  // Inst{127-96} = 0;
+
+  let VTXInst = 1;
+}
+
 class LoadParamFrag <PatFrag load_type> : PatFrag <
   (ops node:$ptr), (load_type node:$ptr),
   [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
@@ -1254,6 +1298,133 @@ def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
   [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
 >;
 
+class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+    : VTX_WORD0_eg, VTX_READ<name, buffer_id, outs, pattern> {
+
+  // Static fields
+  let VC_INST = 0;
+  let FETCH_TYPE = 2;
+  let FETCH_WHOLE_QUAD = 0;
+  let BUFFER_ID = buffer_id;
+  let SRC_REL = 0;
+  // XXX: We can infer this field based on the SRC_GPR.  This would allow us
+  // to store vertex addresses in any channel, not just X.
+  let SRC_SEL_X = 0;
+
+  let Inst{31-0} = Word0;
+}
+
+class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
+                   pattern> {
+
+  let MEGA_FETCH_COUNT = 1;
+  let DST_SEL_X = 0;
+  let DST_SEL_Y = 7;   // Masked
+  let DST_SEL_Z = 7;   // Masked
+  let DST_SEL_W = 7;   // Masked
+  let DATA_FORMAT = 1; // FMT_8
+}
+
+class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
+                    pattern> {
+  let MEGA_FETCH_COUNT = 2;
+  let DST_SEL_X = 0;
+  let DST_SEL_Y = 7;   // Masked
+  let DST_SEL_Z = 7;   // Masked
+  let DST_SEL_W = 7;   // Masked
+  let DATA_FORMAT = 5; // FMT_16
+
+}
+
+class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
+                   pattern> {
+
+  let MEGA_FETCH_COUNT = 4;
+  let DST_SEL_X        = 0;
+  let DST_SEL_Y        = 7;   // Masked
+  let DST_SEL_Z        = 7;   // Masked
+  let DST_SEL_W        = 7;   // Masked
+  let DATA_FORMAT      = 0xD; // COLOR_32
+
+  // This is not really necessary, but there were some GPU hangs that appeared
+  // to be caused by ALU instructions in the next instruction group that wrote
+  // to the $ptr registers of the VTX_READ.
+  // e.g.
+  // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
+  // %T2_X<def> = MOV %ZERO
+  //Adding this constraint prevents this from happening.
+  let Constraints = "$ptr.ptr = $dst";
+}
+
+class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
+                   pattern> {
+
+  let MEGA_FETCH_COUNT = 16;
+  let DST_SEL_X        =  0;
+  let DST_SEL_Y        =  1;
+  let DST_SEL_Z        =  2;
+  let DST_SEL_W        =  3;
+  let DATA_FORMAT      =  0x22; // COLOR_32_32_32_32
+
+  // XXX: Need to force VTX_READ_128 instructions to write to the same register
+  // that holds its buffer address to avoid potential hangs.  We can't use
+  // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
+  // registers are different sizes.
+}
+
+//===----------------------------------------------------------------------===//
+// VTX Read from parameter memory space
+//===----------------------------------------------------------------------===//
+
+def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
+  [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
+>;
+
+def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
+  [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
+>;
+
+def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
+  [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
+>;
+
+def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
+  [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
+>;
+
+//===----------------------------------------------------------------------===//
+// VTX Read from global memory space
+//===----------------------------------------------------------------------===//
+
+// 8-bit reads
+def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
+  [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
+>;
+
+// 32-bit reads
+def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
+  [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
+>;
+
+// 128-bit reads
+def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
+  [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
+>;
+
+//===----------------------------------------------------------------------===//
+// Constant Loads
+// XXX: We are currently storing all constants in the global address space.
+//===----------------------------------------------------------------------===//
+
+def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
+  [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
+>;
+
+
 } // End Predicates = [isEG]
 
 //===----------------------------------------------------------------------===//
@@ -1408,13 +1579,71 @@ let hasSideEffects = 1 in {
     let END_OF_PROGRAM = 1;
   }
 
+} // End Predicates = [isEGorCayman]
+
 //===----------------------------------------------------------------------===//
-// Memory read/write instructions
+// Regist loads and stores - for indirect addressing
 //===----------------------------------------------------------------------===//
 
-class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
-    : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
-      VTX_WORD1_GPR, VTX_WORD0 {
+defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
+
+//===----------------------------------------------------------------------===//
+// Cayman Instructions
+//===----------------------------------------------------------------------===//
+
+let Predicates = [isCayman] in {
+
+let isVector = 1 in {
+
+def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
+
+def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
+def MULHI_INT_cm : MULHI_INT_Common<0x90>;
+def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
+def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
+def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
+def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
+def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
+def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
+def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
+def SIN_cm : SIN_Common<0x8D>;
+def COS_cm : COS_Common<0x8E>;
+} // End isVector = 1
+
+def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
+def : SIN_PAT <SIN_cm>;
+def : COS_PAT <COS_cm>;
+
+defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
+
+// RECIP_UINT emulation for Cayman
+// The multiplication scales from [0,1] to the unsigned integer range
+def : Pat <
+  (AMDGPUurecip i32:$src0),
+  (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
+                            (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
+>;
+
+  def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
+    let ADDR = 0;
+    let POP_COUNT = 0;
+    let COUNT = 0;
+  }
+
+def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
+
+
+def RAT_STORE_DWORD_cm : EG_CF_RAT <
+  0x57, 0x14, 0x1, (outs),
+  (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr),
+  "EXPORT_RAT_INST_STORE_DWORD $rw_gpr, $index_gpr",
+  [(global_store i32:$rw_gpr, i32:$index_gpr)]
+> {
+  let eop = 0; // This bit is not used on Cayman.
+}
+
+class VTX_READ_cm <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+    : VTX_WORD0_cm, VTX_READ<name, buffer_id, outs, pattern> {
 
   // Static fields
   let VC_INST = 0;
@@ -1425,53 +1654,18 @@ class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
   // XXX: We can infer this field based on the SRC_GPR.  This would allow us
   // to store vertex addresses in any channel, not just X.
   let SRC_SEL_X = 0;
-  let DST_REL = 0;
-  // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
-  // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
-  // however, based on my testing if USE_CONST_FIELDS is set, then all
-  // these fields need to be set to 0.
-  let USE_CONST_FIELDS = 0;
-  let NUM_FORMAT_ALL = 1;
-  let FORMAT_COMP_ALL = 0;
-  let SRF_MODE_ALL = 0;
+  let SRC_SEL_Y = 0;
+  let STRUCTURED_READ = 0;
+  let LDS_REQ = 0;
+  let COALESCED_READ = 0;
 
   let Inst{31-0} = Word0;
-  let Inst{63-32} = Word1;
-  // LLVM can only encode 64-bit instructions, so these fields are manually
-  // encoded in R600CodeEmitter
-  //
-  // bits<16> OFFSET;
-  // bits<2>  ENDIAN_SWAP = 0;
-  // bits<1>  CONST_BUF_NO_STRIDE = 0;
-  // bits<1>  MEGA_FETCH = 0;
-  // bits<1>  ALT_CONST = 0;
-  // bits<2>  BUFFER_INDEX_MODE = 0;
-
-
-
-  // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
-  // is done in R600CodeEmitter
-  //
-  // Inst{79-64} = OFFSET;
-  // Inst{81-80} = ENDIAN_SWAP;
-  // Inst{82}    = CONST_BUF_NO_STRIDE;
-  // Inst{83}    = MEGA_FETCH;
-  // Inst{84}    = ALT_CONST;
-  // Inst{86-85} = BUFFER_INDEX_MODE;
-  // Inst{95-86} = 0; Reserved
-
-  // VTX_WORD3 (Padding)
-  //
-  // Inst{127-96} = 0;
-
-  let VTXInst = 1;
 }
 
-class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
-    : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
+class VTX_READ_8_cm <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_cm <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
                    pattern> {
 
-  let MEGA_FETCH_COUNT = 1;
   let DST_SEL_X = 0;
   let DST_SEL_Y = 7;   // Masked
   let DST_SEL_Z = 7;   // Masked
@@ -1479,10 +1673,9 @@ class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
   let DATA_FORMAT = 1; // FMT_8
 }
 
-class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
-    : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
+class VTX_READ_16_cm <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_cm <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
                     pattern> {
-  let MEGA_FETCH_COUNT = 2;
   let DST_SEL_X = 0;
   let DST_SEL_Y = 7;   // Masked
   let DST_SEL_Z = 7;   // Masked
@@ -1491,11 +1684,10 @@ class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
 
 }
 
-class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
-    : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
-                   pattern> {
+class VTX_READ_32_cm <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_cm <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
+                pattern> {
 
-  let MEGA_FETCH_COUNT = 4;
   let DST_SEL_X        = 0;
   let DST_SEL_Y        = 7;   // Masked
   let DST_SEL_Z        = 7;   // Masked
@@ -1512,11 +1704,10 @@ class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
   let Constraints = "$ptr.ptr = $dst";
 }
 
-class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
-    : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
+class VTX_READ_128_cm <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_cm <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
                    pattern> {
 
-  let MEGA_FETCH_COUNT = 16;
   let DST_SEL_X        =  0;
   let DST_SEL_Y        =  1;
   let DST_SEL_Z        =  2;
@@ -1532,20 +1723,19 @@ class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
 //===----------------------------------------------------------------------===//
 // VTX Read from parameter memory space
 //===----------------------------------------------------------------------===//
-
-def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
+def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0,
   [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
 >;
 
-def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
+def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0,
   [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
 >;
 
-def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
+def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0,
   [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
 >;
 
-def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
+def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0,
   [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
 >;
 
@@ -1554,92 +1744,20 @@ def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
 //===----------------------------------------------------------------------===//
 
 // 8-bit reads
-def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
+def VTX_READ_GLOBAL_8_cm : VTX_READ_8_cm <1,
   [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
 >;
 
 // 32-bit reads
-def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
+def VTX_READ_GLOBAL_32_cm : VTX_READ_32_cm <1,
   [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
 >;
 
 // 128-bit reads
-def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
+def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1,
   [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
 >;
 
-//===----------------------------------------------------------------------===//
-// Constant Loads
-// XXX: We are currently storing all constants in the global address space.
-//===----------------------------------------------------------------------===//
-
-def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
-  [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
->;
-
-}
-
-//===----------------------------------------------------------------------===//
-// Regist loads and stores - for indirect addressing
-//===----------------------------------------------------------------------===//
-
-defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
-
-//===----------------------------------------------------------------------===//
-// Cayman Instructions
-//===----------------------------------------------------------------------===//
-
-let Predicates = [isCayman] in {
-
-let isVector = 1 in {
-
-def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
-
-def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
-def MULHI_INT_cm : MULHI_INT_Common<0x90>;
-def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
-def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
-def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
-def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
-def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
-def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
-def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
-def SIN_cm : SIN_Common<0x8D>;
-def COS_cm : COS_Common<0x8E>;
-} // End isVector = 1
-
-def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
-def : SIN_PAT <SIN_cm>;
-def : COS_PAT <COS_cm>;
-
-defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
-
-// RECIP_UINT emulation for Cayman
-// The multiplication scales from [0,1] to the unsigned integer range
-def : Pat <
-  (AMDGPUurecip i32:$src0),
-  (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
-                            (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
->;
-
-  def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
-    let ADDR = 0;
-    let POP_COUNT = 0;
-    let COUNT = 0;
-  }
-
-def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
-
-
-def RAT_STORE_DWORD_cm : EG_CF_RAT <
-  0x57, 0x14, 0x1, (outs),
-  (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr),
-  "EXPORT_RAT_INST_STORE_DWORD $rw_gpr, $index_gpr",
-  [(global_store i32:$rw_gpr, i32:$index_gpr)]
-> {
-  let eop = 0; // This bit is not used on Cayman.
-}
-
 } // End isCayman
 
 //===----------------------------------------------------------------------===//
@@ -1760,7 +1878,7 @@ def CONST_COPY : Instruction {
 def TEX_VTX_CONSTBUF :
   InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
       [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
-  VTX_WORD1_GPR, VTX_WORD0 {
+  VTX_WORD1_GPR, VTX_WORD0_eg {
 
   let VC_INST = 0;
   let FETCH_TYPE = 2;
@@ -1814,7 +1932,7 @@ def TEX_VTX_CONSTBUF :
 def TEX_VTX_TEXBUF:
   InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
       [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
-VTX_WORD1_GPR, VTX_WORD0 {
+VTX_WORD1_GPR, VTX_WORD0_eg {
 
 let VC_INST = 0;
 let FETCH_TYPE = 2;